Module auto_improve 

Autonomous multi-hour lab loop: measure → search configs → re-measure → lock-in.

This is not an LLM rewriting kernels. It is a disciplined experimental program that:

1. Explores a discrete configuration space (resident, coop, kv, mode, backend)
2. Optionally re-samples the best configs (self-improvement via evidence)
3. Tracks plateau / wall-clock budget
4. Emits a lock-in package: best config, full CSV, methodology text, apply script

Plan: docs/plans/inference-superiority-lab-and-toolset-plan.md L4.

Structs

AutoImproveConfig

LabConfig

One point in the searchable config space.

LockInPackage

TrialResult

Functions

default_search_space

Search space for recursive improvement (expandable).

format_lockin_summary

run_auto_improve

Run the recursive lab program; write lock-in package under cfg.out_dir.

try_ollama_decode_tok_s

Probe Ollama decode tok/s if server is up (optional; not a product dependency).